A Bimodal Graph Neural Network with Transfer Learning and Contrastive Learning for Protein-Protein Interaction Site Prediction.

Accurate prediction of protein-protein interaction sites (PPIS) is crucial for understanding molecular mechanisms and advancing precision medicine. However, due to the complexity of protein structure and function, as well as the sequential and spatial heterogeneity of PPIS, accurately predicting these sites remains a major challenge. To address the limitations of existing methods in feature representation and class imbalance, this paper proposes TransPPIS, a transfer learning and contrastive learning based method for predicting PPIS using bimodal graphs. Specifically, TransPPIS constructs a structural graph from the three-dimensional structure of a protein and models its sequence as a de Bruijn graph. Then, an unsupervised pretraining combined with a transfer learning strategy is employed, transferring a graph convolutional network pre-trained on a large-scale protein dataset to the PPIS prediction task. In addition, we design a supervised contrastive learning module, which optimizes cross-modal feature consistency and class discriminability through a dual mechanism of collaborative contrast and positive-negative sample contrast, effectively mitigating class imbalance. Experimental results demonstrate that TransPPIS consistently outperforms state-of-the-art methods on multiple independent test sets, exhibiting superior prediction accuracy, robustness, and generalization. Ablation studies further confirm the effectiveness of each module and validate the synergistic benefits of structured modeling, transfer learning, and supervised contrastive learning. These results suggest that TransPPIS provides a powerful tool for efficiently and accurately deciphering complex protein interaction networks. The source code of TransPPIS is available at https://github.com/houyichi6-netizen/transppis .